Runtime health
- HTTP endpoints and expected status
- System services and scheduled jobs
- Disk pressure and host reachability
- Application-specific health checks
Proactive monitoring · evidence before noise
She catches quiet failures before they get loud.
Ophelia watches the boring operational facts continuously, separates a live failure from an old warning, and escalates through the least disruptive channel that will actually reach you.
What she watches
The failure surfaces that dashboards forget to connect.
Every check is narrow, observable, and configured for the system it protects. Ophelia does not infer “down” from an empty response when the real answer is “I could not reach it.”
Operational continuity
- Backup age, size, and off-site evidence
- Credential health from redacted summaries
- Messaging and voice capacity
- Daily and monthly model spend ceilings
Connected work signals
- Important inbox activity
- Upcoming calendar events
- Read-only pipeline movement
- Mode-aware work and personal routing
Signal path
Observe once. Verify. Escalate only as far as needed.
Observe
A purpose-built check returns a fact, severity, timestamp, and redacted evidence.
Deduplicate
The same unresolved condition does not become a new emergency every polling cycle.
Route
Routine items wait for the brief. Warnings reach chat. Critical issues can move to SMS or voice.
Resolve
Current health and acknowledged history stay distinct, so an old incident cannot masquerade as a live outage.
Noise control
Silence is useful when it is deliberate.
Ophelia respects quiet hours and operating mode. A work-only pipeline notice can wait until work mode; a genuinely critical failure can climb from Telegram to SMS to a call. Each escalation is driven by severity and delivery state, not drama.
Briefs over chatter
Routine facts are collected into a morning or afternoon rundown instead of arriving as scattered pings.
Feedback retunes the alert
Reply to a recent warning in plain language to mute or correct it. The adjustment is explicit and reversible.
Degraded truth stays truthful
If a language model is unavailable, the verified fact still ships in plain wording. Style may degrade; evidence does not.
Security model
Monitor credentials without turning alerts into leaks.
Credential checks consume redacted health summaries—presence, expiry, and logged-out state—not reusable secret material. Outbound messages pass through a final scrubber, and private panels remain loopback-only.
Availability
Current coverage, connected betas, and the next boundary.
Core infrastructure watch
Endpoints, system services, storage, scheduled jobs, backup freshness, redacted credential health, communications quota, and model spend.
Calendar, inbox, and pipeline context
These checks are implemented for configured accounts and require deliberate authorization, identity verification, and per-tenant routing.
Fleet-aware monitoring
Private-network visibility across more owner devices is planned after Tailscale enrollment and machine-level access controls are verified.